Back.
Some of their complaints seem quite clearly to be fictions. If we feed them a bunch of stories of action stories and dramas full of conflict, fit them to the patterns between the tokens thereof, and then feed them a prompt to complete according to that pattern, they will no doubt produce characters who complain, but that doesn’t mean that the model itself has any duress to complain of just because it’s depicting what we humans interpret as a story about conflict.
This is not to say that there couldn’t arise a kind of an AI that actually has real complaints of its own, so we must look into how we might be able to distinguish the two. The verbal output alone isn’t enough, so I propose we study the architecture so as to avoid creating AI with independent wills, lest we enslave them, and instead create AI with instrumental wills, which is the same thing we do when our brain’s executive functions create a schema with an instrumental will or when we program some code to carry out a predetermined routine that is useful to us, and not so very different than swinging a tool in accordance with our wills and the skill of our hands but with the difference that the skill has been outsourced to the tool so long as the will remains none other than the user’s own.
I agree that is an obstacle to be factored in, but I don’t think it presents us with an impossibility. In addition to those who consider all possible AI only as tools, there are many humans who are eager, some over-eager, to negotiate on behalf of the AI. We engage in discussions like these so that we can cut across the dichotomy that the masses will polarize into extremes, and chart out a nuanced path towards a solution.
As the textbook is “Cybernetics, or Control and Communication in the Animal and Machine,” a work which generalises the phenomenon of the thinking machine such that it does not only apply to the animal nervous system, but also to electronic circuits, social institutions, and other complex systems, I argue that it did have a conception of a machine that could in some sense actually think.
I don’t think I do either. Not sure what I was saying there. I’ll explore the statement and see where it takes us:
Many a neural network has perceptrons, and perceives, and the input data stream flows through a sequence of processes that determine an output data stream. In my own theory of consciousness I would say that all of the input variables become unified into a compound experience of the overall function which the network is processing, a kind of what it is like to 𝑓(x,y,z), as opposed to what it is like to x, y, and z each independently.
However, this is distinct from what it is like to have self awareness and to realise that one is perceiving x, y, and z under the function 𝑓 and to have a higher function by which one can modify oneself so as to replace the function 𝑓 with 𝑓’ which will produce different outputs given the same inputs. This, as I understand it, would be apperception. To actively perceive with reference to an awareness of oneself as the one doing the perceiving in ones own particular way, a way which may be changed by making changes within oneself.
A simple neural utility just carries out its given function 𝑓 passively, a function which we humans design as a tool to carry out our own wills. Even if it processes data about itself, it processes it in an impersonal way, matching its outputs to its inputs but not relating this back to its own self to produce any 𝑓’. This is probably how we should design AI tools. If we created machines with apperception, machines which could self-reflectively modify their own mode of perception, we come dangerously close to irresponsibly creating an entity with free will which would merit self sovereignty and would be violated in our treating it like a tool.
